RPA: The Best Distribution Channel for Enterprise AI
Editor’s note: the following is an excerpt from an interview David Eddy conducted with Prabhdeep (PD) Singh, Head of Artificial Intelligence (AI) Products at UiPath. This is a guest post by Eddy and is the first of a two-part series.
Eddy: PD, you’ve been Head of AI Products at UiPath since joining the company three years ago. How would you describe the background you brought to that role?
Singh: I started my career—and we’re talking about the early 2000s, when Google transformed the internet search space by making knowledge accessible—developing simple content recommendation systems and knowledge management systems. I wound up building several startups around matching products to consumers and clustering those product categories together. They were something similar to Pinterest but way before its time.
At that time, and today as well, I’ve always wanted my work to have a real impact on the world we live in, and that pulled me to the healthcare field. More specifically, healthcare data, and it’s never lost its appeal. Just recently, I tweeted about AI in healthcare: how it can leverage EMR data to identify high-risk patients, detect lesions, and even use X-ray images to automate pneumonia detection.
Back in 2006, when I was with Microsoft, resources to pursue healthcare innovation were available, so I began working with a group of doctors to create a new data warehouse. Named Amalga, it aggregated patient data from all different healthcare sources. My vision was to put machine learning on top of that data and position Amalga as a unified health enterprise platform. It was the Snowflake for healthcare data in 2010-2011. However, by 2013, Steve Ballmer was the CEO at Microsoft and he decided Amalga wasn’t a core technology, so it was spun off as a joint venture with GE.
After the spinoff, my passion for helping people with machine learning [ML] led me to Microsoft Research, where I began collaborating with the Office Group. In 2014, our first success was Office Lens (now called Microsoft Lens), a mobile app that could do basic OCR [optical character recognition], image correction, and then store it in the cloud. Next, we transformed Outlook Search by implementing deep learning.
Eddy: It must have been an exciting time for you.
Singh: Yes, in many ways it was. But, one question kept bothering me—can AI/ML really make a significant impact to a company or line of business? That was a hard question for me to answer six years ago. Frankly, a great number of business people find it a challenge to answer today.
To find the answer, I spun off a new initiative: Sales Intelligence, Business AI, with a charter to transform Microsoft sales and marketing. Our group brought reinforcement learning and deep learning to the sales funnel and across inside sales and field sales. It was also applied to operations. For example, lead and churn management.
Eddy: With that level of success at Microsoft, what attracted you to UiPath?
Singh: Yes, the outcomes at Microsoft were certainly OK. But here’s what bothered me: it took us too long to build those AI systems. Two and a half years, with one of the best teams in the world, a team that’s built most of the best state-of-the-art NLP [natural language processing] models in the world for question answering and sentence completion.
I found myself thinking, “If a team like this takes two and a half years to build a basic ML/AI model for sales and marketing prediction, how will this technology ever make an impact in companies that don’t have those people?”
I’d also seen first-hand that data scientists were only spending 10 to 20% of their time on modeling. Ancillary functions like collecting data or integrating with the systems people were using, took up all the rest. So, when I looked at UiPath, I realized the company’s technology could really accelerate the entire ML/AI process and impact companies in a big way.
Then, after I joined the company, my vision for the AI/ML team I was building was profoundly influenced by a chat I had with Laela Sturdy, UiPath board member and CapitalG partner. Her point was RPA should be the main distribution channel for AI.
When you think about that for a moment, you realize it’s a pretty profound observation. AI isn’t able to distribute itself to end users. It needs a channel, and RPA should be it. So, everything we’ve built has been guided by the basic principle that RPA should be a primary vehicle for delivering AI.
Eddy: Are you talking about delivering AI via general RPA functionality or specific product features and capabilities?
Singh: Both, and more. For example, the industry-leading functionality of our Computer Vision uses a range of ML drivers for perceiving screen elements, controls, and OCR enablement. At the same time, our team has baked AI into almost every aspect of UiPath Platform products, including Process Mining, Task Mining, Studio, and Orchestrator.
Eddy: How do you explain the synergistic value of RPA and AI/ML to current and prospective customers? What’s the best way you’ve found to get across the impact those technologies can bring to their business?
Singh: Some customers are very traditionally minded in an, “I know my business well, better than any of you—and I don’t need ML.”
In those cases, we talk about technology as waves of disruption. For example, what Amazon has done to classic retailers like Sears, or how Uber and Lyft decimated taxi medallion values by marrying mobile with cabs. Nobody is immune to that risk.
We encourage customers to view AI/ML as an assistive technology and to think of ways it could lead to competitive advantages. If you’re running a logistics company, you might say, “OkK, can AI/ML tell me when specific units in my fleet should go in for proactive maintenance to reduce the fleet downtime?”
In fact, predictive maintenance for really expensive machines like power station turbines and aircraft engines was one the first applications of AI.
Eddy: Once you’ve got the customer engaged, what expectations for UiPath capabilities do you set for them?
Singh: The takeaway we want our customers to have is this: if you ever see a need for AI/ML in your automated workflows, or workflows you’re thinking of making smarter—UiPath should be the company you think of and talk to. And it doesn’t matter what stage you’re at. We have a very flexible platform that will meet our customers where they are.
Perhaps you’ve got an internal data science team building models and finding “levers of acceleration,” as I call them, within your business processes. UiPath AI Center will take care of managing, scaling, and retraining those models, allowing your RPA developers to easily integrate and use them in any number of ways. If you don’t have that team, we have out-of-the-box ML models for immediate processing of invoices, insurance claims, bill-of-ladings, and many, many more.
Eddy: What do you say to skeptics that point to articles like MIT’s <em>The Way We Train AI is Fundamentally Flawed</em>, which says no one can be certain if most ML models will actually work in the real world?
Singh: The problem discussed in the MIT article is underspecification. What it means is this: machine learning models may test out equally well after development but not perform as well in real-world use. Some will do better, others worse, and there’s no way to predict those outcomes.
It is a serious issue, and how we protect UiPath customers from it speaks to what really differentiates us from our competitors. We don’t claim to have the best-in-class out-of-the-box models. We are not one of those OCR or RPA competitors who say, ‘Oh, you can expect 99% accuracy with our OCR technology.’ That’s not the business we are in.
We’re in the business of saying—and it comes from the book Mindset written by Stanford professor Carol Dweck—’we are not going to be the all-knowing company, but we will be the all-learning company, we will learn faster than anyone else.’